Benchmarking Five Machine Learning Models for Accurate Steel Plate Defect Detection

Abstract

Early detection of defects in steel plates is essential to ensure structural integrity and product quality in the metal manufacturing industry. This study compares the performance of five machine learning algorithms Support Vector Classifier (SVC), Nu-Support Vector Classifier (NuSVC), Decision Tree (DT), Random Forest (RF), and CatBoost (CB) to classify seven categories of steel plate defects using 26 technical features from a publicly available dataset on Kaggle. The preprocessing pipeline included outlier detection (IQR method), class imbalance correction using SMOTE, and feature normalization via StandardScaler. The models were evaluated using classification metrics such as Accuracy, Precision, Recall, F1-Score, ROC-AUC, and Log Loss. Results revealed that the CatBoost algorithm achieved the most balanced and consistent performance, with an AUC of 0.93, accuracy of 68.3%, and the lowest Log Loss value (0.786). In contrast, the Decision Tree showed severe overfitting with perfect training performance but poor generalization (Log Loss = 15.72). This study highlights the promise of CatBoost as an interpretable and efficient solution for automated defect detection in steel manufacturing, while also offering transparent reproducibility pathways for further research.